Development of a Preliminary Diagnostic Tool for the Segmentation of Benign Jaw Lesions in CBCT Images Using nnU-Net v2: An Artificial Intelligence-Based Approach.
Authors
Affiliations (7)
Affiliations (7)
- Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Selçuk University, Konya, 42130, Turkey.
- Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Alanya Alaaddin Keykubat University, Antalya, 07425, Turkey.
- Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Kocaeli University, Kocaeli, 41190, Turkey. [email protected].
- Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Ankara University, Ankara, 06560, Turkey.
- Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Eskisehir Osmangazi University, Eskisehir, 26040, Turkey.
- Department of Mathematics-Computer, Eskisehir Osmangazi University Faculty of Science, Eskisehir, 26040, Turkey.
- Center of Research and Application for Computer-Aided Diagnosis and Treatment in Health (ESOGU-SABIT), Eskisehir Osmangazi University, Eskisehir, Turkey.
Abstract
Accurate segmentation of jaw lesions on cone beam computed tomography (CBCT) images is essential for diagnosis and treatment planning. While previous studies have investigated automated segmentation, most were limited to 2D data, and the performance of newer architectures like nnU-Net v2 remains unexamined. This study evaluated the nnU-Net v2 algorithm for 3D segmentation of jaw lesions and compared its speed to manual expert segmentation. Due to significant class imbalance in the dataset, training a single multiclass segmentation model could have led to dominance of the majority class and reduced Dice similarity coefficient (DSC) for minority classes. Therefore, three separate models were developed: the first for all benign jaw lesions (355 CBCT), the second for radicular cysts (305 CBCT), and the third for dentigerous cysts (33 CBCT). The radicular and dentigerous cyst datasets consisted of the same images used in the initial model but were additionally utilized as subsets for lesion-specific evaluation. The models were assessed using precision, recall, and DSC, compared with expert manual segmentation time, and tested on an external dataset. DSC values were 0.70 ± 0.08, 0.70 ± 0.21, and 0.72 ± 0.04 for the benign, radicular, and dentigerous cyst models, respectively. Automated segmentation required less than 1 min, approximately 15 times faster than manual segmentation. DSC on the external dataset ranged from 0.84 to 0.87. The nnU-Net v2-based deep learning models demonstrated consistent segmentation performance across both general and lesion-specific datasets. Furthermore, these models can automatically segment lesions in seconds, providing a substantial time advantage over manual segmentation by expert clinicians.